Prediction model of malignant risk of gastrointestinal stromal tumor based on endoscopic ultrasonography

Introduction: Gastrointestinal stromal tumors (GIST) are clinically heterogeneous mesenchymal tumors, and accurate assessment of their malignancy is crucial for determining treatment strategies and predicting patient prognosis. This study aims to establish and validate a preoperative prediction model for the malignant risk of GIST based on endoscopic ultrasound (EUS) findings. Methods: In this retrospective study, we collected data from patients with GIST at Sichuan Provincial People's Hospital. Statistical analyses involved univariate and multivariate logistic regression models, with Firth's bias reduction penalty likelihood logistic regression employed to address data separation and multicollinearity. Model discrimination was assessed using AUC, and the Bootstrap ROC method was used for validation. Finally, a nomogram was constructed to visually represent the predictive model. Results: In this study, we analyzed 500 patients with GIST, including low (n=393) and high-risk groups (n=107). Univariate and multivariate logistic regression analyses revealed hard elastography and contrast-enhanced EUS (CE-EUS) of visible tumor vessels as potent independent risk factors associated with GIST malignancy, Firth's bias reduction penalty likelihood logistic regression confirmed hard elastography (OR=146.222, 95%CI: 12.597-1017.552) and CE-EUS with visible tumor vessels (OR=402.853, 95%CI: 12.732-7633.787) as significant predictors. Finally, we identified five optimal predictors (including size, shape, echogenicity, elastography, and visible tumor vessels). The predictive model demonstrated exceptional accuracy with an AUC of 0.9998, further validated by Bootstrap ROC analysis showing an AUC of 1. Conclusion: A nomogram incorporating tumor size, shape, echogenicity, elastography, and visible tumor vessels was developed. However, because internal validation indicated overfitting, the model’s clinical utility requires rigorous external validation.